15 collaborators online

Team projects. Real contributions. Two weeks.

Upskill, contribute, and build evidence that gets you hired, all while learning about the field.

The upskilling gap

Reading about AI safety isn't the same as doing the work.

Thousands of people finish an AI safety intro course every year. The fellowships that pick up where the courses leave off accept roughly 5%. The other 95 stall at the same place: motivated, conceptually trained, no structured way to upskill into actual contribution.

0+
finish an AI safety intro course each year
Motivated, conceptually trained, looking for what to do with it next.
0%
find a fellowship-style next step where they can upskill
Generator, MATS, Astra, Anthropic Fellows accept 2 to 5% of applicants. The other 95 have no shaped path.
0 days
from joining Mangrove to shipping real work
A team, a problem, peers who rate your craft, an artifact you can point to. The bridge from interest to contribution.

The field needs more contributors. The pipeline to make them barely exists. Mangrove is the missing layer between intro courses and the work itself. Real projects, real teams, real artifacts, in two weeks.

How it works

Project sprints. From ideas to artifacts.

Mangrove is always on. No application deadline. No fixed start window. A team forms when the conditions to do good work are met, and not before.

Step 01 of 06

Paint when you're available.

Real teams need real overlap. The platform respects your timezone and your life.

01A 24 × 7 weekly grid. Toggle the hours you'd be available for recurring sprint work, in your local timezone.
02Stored in UTC. Overlap-matching crosses timezones without anyone having to think about it.
03Re-editable any time. Schedule shifts; the system honors the new grid on the next match.
Async teams fall apart when no two people are awake at the same time. The grid is the contract that prevents that, before anyone gets staffed onto anything.
Your availability · local timezone 12 hrs/week
MonTueWedThuFriSatSun
8a
9a
10a
11a
12p
1p
2p
3p
7p
8p
9p
on-call maybe Tap a cell to toggle · saved on change
Step 02 of 06

Pick problems worth your time.

Members and organizations post ideas. Two buttons that mean different things.

01Each idea explains itself. Problem statement, why it matters, what shipping would look like, plus an author-set commitment threshold (min 2, default 3).
02"I'm interested." Engagement only. Follow the conversation, get notified when the author refines the idea. No commitment.
03"I'd join this." Soft commitment. You're saying: if the team forms and the timing works, staff me. The platform tracks both counts separately.
Curiosity and conscription are different signals. Treating them the same is what makes "interested?" forms feel like a trap. We keep them apart on purpose.
Idea pool · overlapping with your availability 47 visible
Map every public alignment grant in 2025 → outcomes registry
field-build  ·  posted by @leah · 5 days ago
threshold
2 / 3
Plain-language briefings of CHAI papers for policy staffers
comms  ·  posted by Redwood (org)
threshold
1 / 3
Step 03 of 06

When the team forms, the team forms.

Threshold reached. Schedules overlap. The team forms itself.

01Two conditions, both required. Enough "I'd join" votes to clear the threshold, and a subgroup of those committers whose availability grids actually overlap.
02The team forms automatically. Committed members get auto-staffed as the seed team. The first meeting is auto-scheduled into the overlap slot.
03Calendar invite lands. Google Meet link + .ics file delivered by email and in-app. Nothing else for you to set up.
No application deadline. No "next cohort opens June 8." The system fires the moment the conditions to do good work are met, whether that's tomorrow, next week, or whenever the overlap shows up.
Team formed · just now
Sleeper-agent eval replication on open weights
Threshold hit (4 of 3) and a 4-person overlap slot was found. You've been auto-staffed as the seed team.
Team
YOU MC PI KS
Kickoff
Tue · 4:00 PM UTC · 9 AM PT · 5 PM CET
Calendar
.ics attached · added to all 4 members' inboxes
Step 04 of 06

Scope it together.

First meeting writes the proposal. Roles, deliverables, what done looks like.

01The team scopes its own work. Mangrove doesn't pick your approach. It just makes sure the team has the same conversation at the same time.
02Roles get assigned. Project Lead, Data Wrangler, Comms Lead, Field Interviewer, whatever the work needs. Every role is named on Day 1.
03Decision log opens. Weekly check-ins get auto-scheduled into the team's shared overlap. Every pivot logged with the why.
Role clarity is the coordination engine for ER teams, film crews, disaster-response units, anywhere strangers do complex work together. You don't argue about who does what; you do the work.
Kickoff agenda · 60 min starts in 12 min
Sleeper-agent eval replication kickoff
Tue · 4:00 PM UTC · 4 members
▶ Join Meet
0:00
Intros + intent. What each member is bringing, what they want out of two weeks
0:10
Scope the v0. The smallest shippable shape of the deliverable, written collaboratively in the kickoff doc
0:30
Roles + deliverables. Project Lead, Data Wrangler, Eval Engineer, Memo Author
0:45
Schedule check-ins. Auto-placed in the team overlap window, Mon + Fri
0:55
Decision log opens. First entry: "we picked v0 over v1 because…"
Step 05 of 06

Two weeks. Real artifacts.

Async-first. Weekly check-ins. Decision log captures every pivot. Action items have owners.

01Single source of truth. One repo. One doc. One channel. The platform threads them together.
02Auditable decision trail. Every pivot logged with the why. Action items tied to check-ins with assignees and due dates.
03Mentor seat optional. A mentor with the right expertise can drop in for one or two check-ins. No long-term contract.
Long enough to scope, build, revise. Short enough to fit around school, jobs, family. Long enough that the artifact matters. Short enough that trying is cheap.
Sprint · Day 7 of 14 in flight
DAY 1 · MON
Kickoff: scope written, roles assigned, decision log opened
DAY 3 · WED
Async work; first commit hits the repo
DAY 5 · FRI
Check-in #1. Scope held; mentor flagged one risk
DAY 7 · SUN
Data layer done; 2 of 4 evals running
DAY 9 · TUE
Check-in #2. Full v0 expected by EOW
DAY 12 · FRI
Draft frozen for peer + mentor review
DAY 14 · SUN
Ship. Memo + repo + invite stakeholders. Debrief meeting auto-scheduled.
Step 06 of 06

Rate the role, not the person.

Role-anchored ratings. The aggregate becomes your per-role skill graph. Your portfolio.

01Each teammate rates each other. Anchored to the specific role they played on this sprint: Data Wrangler, Project Lead, Comms Lead.
02No global star average. The aggregate by (you, role) is the skill graph. A Project Lead score and a Data Wrangler score live separately.
03Public by default. Linkable to a hiring manager, a collaborator, a grantmaker. Evidence is what they actually want to see.
90% of platform ratings on the open web are five stars. Anchoring to role craft, not the person, is how we get signal instead of inflation. The hiring side of the field has been begging for this artifact for years.
Peer review · Sleeper-agent eval sprint 3 / 4 submitted
Mike Chen
DATA_WRANGLER · scraping, python
want again
Priya Iyer
FIELD_INTERVIEWER · comms
want again
Kai Sato
EVAL_ENGINEER · ML, evals
growing
Leah Park
PROJECT_LEAD · PM, writing
want again
Your portfolio

Get hired by employers.

After a few sprints, your profile is a real piece of evidence, not a self-rated bio. Hiring managers, grantmakers, and collaborators can see the crafts you've built reputation in, anchored to projects you actually shipped, plus an honest record of whether you showed up.

Sample member profile
LP
Leah Park
@leah · 6 sprints · joined Feb 2026
6Sprints shipped
4Distinct roles
14Teammates
Commitment record · last 6 projects
5 of 6 shipped · 83% show-up rate
P-0042
Alignment grants → outcomes registry
Project Lead · 5 funders mapped · memo published on LessWrong
★★★★☆
P-0038
Sleeper-agent eval replication on open weights
Data Wrangler · public notebook · 4 weights tested
★★★★★
P-0031
CHAI paper → policy-staffer brief series
Memo Author · 4 briefs published
★★★★☆
P-0025
Cross-lab researcher directory (v0)
Field Interviewer · 23 calls completed
★★★★★
Project Lead writing field-build interviewing python scraping sql + eval design + comms
Mike Chen Priya Iyer Kai Sato +8 others
"
Each chip is the average rating across every teammate who saw her play that role. Strong means consistent high marks. Medium means real but newer. Growing means she's leaning in: a skill she wanted to stretch on, with peers willing to vouch for the trajectory.
Features

What gets you across the finish line.

Group projects fail in predictable ways. Vague feedback. Fading motivation. Coordination overhead. No artifact at the end. Mangrove is built around those failure modes so you can spend your time on the actual work.

Clear feedback on your skills.

After each sprint, your teammates rate the role you actually played: Data Wrangler, Project Lead, Comms Lead. You see which crafts you got high marks for, and which ones you're still growing into. No five-star inflation, no vague "great job."

A team that expects you to show up.

Three teammates and a kickoff on your calendar do more for motivation than any reading list. The work gets done because four people agreed to do it, not because you white-knuckled it alone at 9 PM.

Group projects without the friction.

Auto-scheduled meetings into the team's overlap window. One repo, one doc, one channel, threaded together. A decision log that captures every pivot. The platform handles the coordination so the team can focus on the problem.

A real artifact at the end.

Every sprint ships something public: a paper draft, a prototype, a memo, a dataset. It lives at a URL you can send a hiring manager, a grantmaker, or a collaborator. No demo day theatrics. Just the work.

Sized for normal lives.

Two-week project arc. Async-first, with one or two live check-ins in your team's overlap window. Fits around a job, a degree, a family. Long enough that the artifact matters, short enough that trying is cheap.

Mentor drop-ins on tap.

Senior researchers and operators opt into weekly check-in slots in their area. Your team gets one or two from someone with the right expertise. They flag a risk, point at the right paper, or ask the question that makes you pivot on Day 5 instead of Day 12.

Common questions

What people ask before signing in.

Fellowships are application-based, run on a fixed cohort calendar, and accept 2 to 5% of applicants. Mangrove has no application and no start date. You sign in, paint your availability, and get auto-staffed onto a team the moment an idea hits its threshold and your schedules overlap. BlueDot teaches you the concepts; Mangrove is where you go after to actually do the work.
No. The first wave of sprints is being seeded now. Small, hand-recruited, instrumented. The platform we ship is the prototype we're testing. Every feature is something we believe matters, and the data from these first sprints tells us which features were right. If you join, you're part of how Mangrove becomes Mangrove.
Most projects need a mix of skills: engineering, writing, interviewing, project management, comms. If you've finished an intro course (BlueDot, ARENA) or have real-world skills you want to apply to safety, you have something a team needs. Non-technical contributors lead a meaningful fraction of projects. Policy briefs, field-building maps, comms work, coordination tooling.
That's the design goal, and the part the first wave of sprints is built to test. Your profile shows real artifacts you shipped, role-anchored peer ratings from teammates, and an honest commitment record, the exact evidence hiring managers and grantmakers say they can't get from résumés. We're partnering with labs and field-builders during the first wave of sprints to make sure profiles get reviewed, and to validate that the signal we're producing is the signal they actually want.
Most sprints run roughly 6 to 10 hours a week, for two weeks. The kickoff and one or two check-ins are live in the team's overlap window; the rest is async. The platform won't auto-staff you onto more than you opted in for. You control how much you take on and pause whenever life needs you to.
Roles are assigned at kickoff and the decision log is shared, so it's visible when someone isn't delivering. Peer ratings at the end capture it honestly. People who repeatedly commit and disappear lose access to "I'd join this" until they rebuild trust. Most teams self-correct fast because everyone wants the artifact to ship. The first wave of sprints will tell us whether this holds. If it doesn't, we'll add structure where it broke.
Yes. Availability is the matching mechanism, not geography. You paint your local hours; the platform handles UTC translation and only forms teams whose members actually share live overlap. Teams routinely come together across 3–4 timezones, including Asia-Pacific, Africa, and South America.
Mostly no. It's volunteer/portfolio work, the way Wikipedia and Polymath are. The artifact you ship and the role-anchored ratings are the compensation, in the form of evidence that gets you hired or funded next. Some org-posted problems come with a stipend or grant; that's per-problem and visible on the idea card before you commit.
Yes. "I'm interested" is engagement only. It follows the conversation and signals encouragement to the idea's author, but doesn't commit you to anything. You can lurk indefinitely, then upgrade to "I'd join this" when an idea grabs you and the timing is right.

From caring about AI safety to building it.

Sign in. Paint your availability. See what's in the pool. The next team forms when the conditions are right.

Already a member? try.mangrove.one